Detecting ODE Mells in Java Rojects: A Omparative Tudy of Raditional Achine Earning and EEP Earning Approaches

Abstract

Code smells indicate declining software quality, yet their detection is challenging. While data driven approaches are promising, a large-scale comparison between traditional Machine Learn-ing (ML) and Deep Learning (DL) for this task is lacking. This study empirically compares their effectiveness for detecting class-level code smells from a large, metric based dataset derived from 525 open-source Java projects. We evaluated three ML and two DL models based on predictive accuracy and computational efficiency. Our findings demonstrate that traditional ML models, particularly XGBoost, consistently and sig nificantly outperform DL models. XGBoost delivered the highest accuracy with high efficiency, while DL models were substantially slower and less accurate. In conclusion, for metric-based code smell detection, tree-based ensemble methods like XGBoost represent the state-of-the-art, offering an optimal balance of high accuracy and low computational cost.